bug in optimised strstr

[Jeff Johnston]

Jeff Johnston wrote:
> Eric Blake wrote:
>> Jeff Johnston <jjohnstn <at> redhat.com> writes:
>>
>>  
>>>   max_suffix = SIZE_MAX;
>>>   j = 0;
>>>   k = p = 1;
>>>   while (j + k < needle_len)
>>>     {
>>>       a = CANON_ELEMENT (needle[j + k]);
>>>       b = CANON_ELEMENT (needle[max_suffix + k]);
>>>
>>> it is the line b=....
>>>
>>> It cannot be correct as you are trying to reference SIZE_MAX + 1 the 
>>> first time through the loop.
>>>     
>>
>> But the comments state:
>>
>>   /* Invariants:
>>      0 <= j < NEEDLE_LEN - 1
>>      -1 <= max_suffix{,_rev} < j (treating SIZE_MAX as if it were 
>> signed)
>> ...
>>
>> On cygwin, this works (in other words, I'm not reproducing the 
>> crash).  The intent is that this line is referencing needle[0].  What 
>> type is size_t on your platform, and the value of SIZE_MAX?  Could it 
>> be that there is some type promotion going on, where the result of 
>> SIZE_MAX+1 results in a 64-bit type containing 2**32 instead of 0, as 
>> is required by modulo math since size_t is unsigned?
>>
>>   
> Ok, I got mixed up on what SIZE_MAX was supposed to be.  For x86, 
> stdint.h is being overridden with one in libc/sys/linux/include that 
> has a wrong value for SIZE_MAX (LONG_MAX).  I am rebuilding now.  I 
> can't say what is happening for arm since it should be using the same 
> stdint.h from libc/include.
>
> -- Jeff J.
>
>
Patch works for x86-linux and is checked in.  I don't have an arm system 
to play around with.  Sam, are you running arm-linux (which isn't in 
newlib)?  That would use the code in question.  Otherwise, can you 
provide more details (e.g. run under gdb and print out the various 
information above)?

-- Jeff J.